Research Terms
Civil Engineering Transportation Engineering
Keywords
Big Data Deep Learning Disaster Management Human Mobility Hurricane Evacuation Machine Learning Social Media Traffic Safety Transportation Networks
Industries
Engineering Modeling, Simulation, & Training (MST)
This UCF invention is an artificial intelligence framework that predicts traffic conditions during hurricane evacuations across large transportation networks. Simply put, the system helps transportation and emergency management agencies anticipate where severe congestion will occur before it happens. Unlike conventional traffic prediction tools that struggle during unusual events such as hurricanes, the technology learns from multiple historical evacuations and can generalize to future storms with different paths, intensities, and evacuation patterns. The system dynamically adapts to changes in sensor availability and road network conditions while delivering network-wide traffic forecasts up to six hours ahead, enabling agencies to make proactive decisions that improve evacuation efficiency and public safety.
Technical Details: The invention combines a Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) architecture to model both spatial and temporal traffic behavior. A master graph representing major transportation corridors is constructed from traffic detector locations and roadway connectivity. Unlike traditional approaches that assume a fixed network structure, the invention dynamically updates the graph at every time step based on active traffic detectors and current network conditions.
The model processes real-time traffic variables, historical traffic trends, time-based indicators, hurricane characteristics, evacuation orders, and population evacuation estimates. Spatial relationships between detectors are extracted through graph learning, while temporal patterns are learned through the LSTM component. Trained on evacuation and traffic data from eleven major Florida hurricanes, the framework can forecast traffic volumes for the next one to six hours while maintaining accuracy even when portions of the sensor network become unavailable. The dynamic graph construction methodology is a key innovation that enables robust performance during highly disruptive emergency situations.
This technology family comprises an integrated suite of computer-implemented systems and methods for real-time traffic prediction, incident management, and network-level control in freeway-arterial corridors. The platform ingests heterogeneous detector data, represents the transportation network as a graph, and applies spatio-temporal artificial intelligence models to forecast near-term traffic states under both routine and incident conditions. Building on these predictions, the system enables automated evaluation, ranking, and selection of incident response strategies - ranging from operational control deployment to coordinated rerouting and signal timing adjustments - using consistent, network-level performance measures. Simulation is leveraged offline to enhance model training and robustness but is removed from real-time operations, enabling low-latency, scalable, and defensible decision support for integrated corridor management.
Technical Details
2025-093-01: Traffic State Prediction and Incident Management with GCN-LSTM Models
This UCF invention provides a real-time operational platform that continuously ingests live detector, incident, weather, and contextual data to predict short-horizon traffic states across freeway-arterial networks. The system represents the corridor as a graph and applies a spatio-temporal predictive model to forecast traffic conditions without executing traffic simulation at run time. Predicted states are used to support immediate incident management actions, and selected strategies can be deployed to traffic field devices, enabling closed-loop, AI-driven traffic operations.
2025-093-02: Synthetic Data Generation for Traffic Models
This UCF invention focuses on robust learning and comparison of incident response strategies by combining real detector data with large libraries of synthetically generated incident scenarios. A graph-based spatio-temporal model is trained offline using simulation-derived scenarios spanning multiple incident locations, durations, severities, and demand levels. During operations, the trained model predicts network traffic states for candidate strategies and evaluates them using network performance measures, enabling scalable, simulation-free strategy selection even for rare or extreme incident conditions.
2025-093-03: Incident-Conditioned Graph Learning for Network-Level Traffic Control
This UCF invention extends predictive traffic modeling by explicitly conditioning forecasts on coordinated control actions, including rerouting options, diversion percentages, and traffic signal timing plan changes. Incident attributes and control decisions are encoded as first-class model inputs, allowing the system to predict how alternative control combinations will affect network-level performance. Strategies are enumerated, evaluated using a uniform performance metric (e.g., total delay), and ranked to support auditable, data-driven control decisions for integrated freeway-arterial incident management.
